A VLM That Learns Space Through Interaction: A Deep Dive into Spatial-Interactor
A walkthrough of Spatial-Interactor, a paper that teaches vision-language models to reason about space not from still images but from "observe–act–observe" interaction trajectories, organized into a three-stage curriculum.
Realtime Full-Duplex AI, Realtime-Venus: A Conversational System That Listens, Watches, and Delegates Work
An abstract-level deep dive into Realtime-Venus, a system that handles real-time full-duplex conversation and background task delegation using one of two 9B model variants per session — the speech-and-video Omni and the audio-only Audio.
Full Participation, Light Compute: How IntBMoE Resolves the Three-Way Dilemma of Mixture-of-Experts
Mixture-of-Experts could never tune participation, execution, and materialization independently. IntBMoE decouples all three with block-level conditioning, and lifted UVCTR by a relative 2.4% in AMap's recommendation service.
The Sharpening Tax: What RL Post-Training Takes Away From AI Agents
Reinforcement learning post-training raises single-shot accuracy, but it can actually narrow the range of tasks a model solves once you give it enough attempts. This paper quantifies that loss as a "sharpening tax" and reduces it with PTGS, a sampler that tunes temperature by difficulty.
ProgramDistill: A Coding-Agent Benchmark That Learns From Apps That Already Work
We introduce ProgramDistill, a coding-agent benchmark that describes a feature's purpose and workflow but makes the agent infer the concrete behavior by interacting with a working reference app and then implement it.
Grading the Roads You Never Took: Dream-RSI
Dream-RSI rewinds the record of past exploration and replays even untried strategies on top of it, scoring them for free.
How a Tiny 7B Takes On Giant Models: ZGCM-1, a Fully Open Foundation Model
ZGCM-1, a fully open 7B model, pairs internal reasoning with external tool use to rival models tens of times its size on math and agentic search. It features a 256K context, FP8 Muon optimization, and agent-assisted R&D.
Gander: A Full-Duplex AI That Talks and Works Without Breaking Stride
Gander, built by researchers at Tencent and Zhejiang University, pairs a real-time conversation model (the Cerebellum) with a separate reasoning backend (the Brain) to create a full-duplex agent that can cut in mid-sentence and ask follow-up questions on the fly.
Dispatch-Style AI PMs as SaaS — How Big Tech's FDE Model Differs from ONTLE
Modeled on Palantir's FDE, Big Tech's dispatch-style AI efforts arrived one after another between May and July 2026. We break down where ONTLE fits among them, by type.
Measuring an Agent's "Taste": A Deep Dive into The Tasteful Agent
This paper defines an agent's ability to make the right choices at decision forks during long-horizon tasks as "taste," and builds an automatic benchmark (Taste-Bench) to measure and train it. Here's a plain-language walkthrough.
WorldCrafter: Building a Consistent Video World Model with Camera-Queryable 3D Memory
WorldCrafter는 시점(카메라)으로 과거 관측을 압축·질의하는 "암묵적 3D 메모리"로 장기 일관성과 카메라 제어 정확도를 끌어올리는 비디오 월드 모델이다.
SoL-Pi: Self-Improving the Agent Harness to Nearly Halve Token Usage
A study that iteratively improves the "harness" layer wrapping a coding agent through an automated research loop, trading a roughly 5–6% drop in score for a 44.7–49.0% cut in tokens and about a one-third reduction in API cost.